Machine and deep learning structural vessel analysis of ruptured and unruptured brain arteriovenous malformations

Abstract

Background

Brain arteriovenous malformations (bAVMs) are complex vascular disorders associated with a considerable lifetime risk of intracerebral hemorrhage. Predicting the likelihood of rupture remains a clinical challenge due to the complex interplay of molecular, anatomical and hemodynamic factors.

Objective

This work proposes an image analysis using machine and deep learning algorithms to evaluate bAVM structural vessel characteristics that may be associated with rupture.

Methods

We trained four artificial intelligence models: isolation forest (IForest), one class support vector machine (OCSVM), variational autoencoder (VAE), and Score Based Models (SBM) to automatically analyze vascular segments of 9 previously ruptured and 11 unruptured bAVMs and detect curve characteristics that may be associated with rupture.

Results

Artificial intelligence models were able to detect curve characteristics inside bAVM that clearly differentiated the ruptured from the unruptured group. The best performing model detecting the rupture risk curves (RRC) was SBM, achieving 98.9% F1-score and 100% Area Under the Curve (AUC). The ruptured group showed significant curves with higher median length and tortuosity, and lower median radius.

Conclusions

The proposed curve analysis approach allowed identifying vessel segments of bAVMs associated with previous rupture with high accuracy. It has the potential to be used as an indicator of hemorrhage and a predictor for rupture in further studies.